600 likes | 1.32k Vues
Rigid-Body Alignment. 3D Scan Matching and Registration, Part II ICCV 2005 Short Course Szymon Rusinkiewicz Princeton University. Pairwise Rigid Registration Goal. Align two partially- overlapping meshes given initial guess for relative transform. Outline. ICP: Iterative Closest Points
E N D
Rigid-Body Alignment 3D Scan Matching and Registration, Part II ICCV 2005 Short Course Szymon Rusinkiewicz Princeton University
Pairwise Rigid Registration Goal Align two partially-overlapping meshesgiven initial guessfor relative transform
Outline • ICP: Iterative Closest Points • Classification of ICP variants • Faster alignment • Better robustness • ICP as function minimization
Aligning 3D Data If correct correspondences are known, can find correct relative rotation/translation
Aligning 3D Data • How to find correspondences: User input? Feature detection? Signatures? • Alternative: assume closest points correspond
Aligning 3D Data • … and iterate to find alignment • Iterative Closest Points (ICP) [Besl & McKay 92] • Converges if starting position “close enough“
Basic ICP • Select e.g. 1000 random points • Match each to closest point on other scan,using data structure such as k-d tree • Reject pairs with distance > k times median • Construct error function: • Minimize (closed form solution in [Horn 87])
ICP Variants Variants on the following stages of ICPhave been proposed: • Selecting source points (from one or both meshes) • Matching to points in the other mesh • Weighting the correspondences • Rejecting certain (outlier) point pairs • Assigning an error metric to the current transform • Minimizing the error metric w.r.t. transformation
Performance of Variants • Can analyze various aspects of performance: • Speed • Stability • Tolerance of noise and/or outliers • Maximum initial misalignment • Comparisons of many variants in[Rusinkiewicz & Levoy, 3DIM 2001]
ICP Variants • Selecting source points (from one or both meshes) • Matching to points in the other mesh • Weighting the correspondences • Rejecting certain (outlier) point pairs • Assigning an error metric to the current transform • Minimizing the error metric w.r.t. transformation
Point-to-Plane Error Metric Using point-to-plane distance instead of point-to-point lets flat regions slide along each other [Chen & Medioni 91]
Point-to-Plane Error Metric • Error function:where R is a rotation matrix, t is translation vector • Linearize (i.e. assume that sin , cos 1): • Result: overconstrained linear system
Point-to-Plane Error Metric • Overconstrained linear system • Solve using least squares
Improving ICP Stability • Closest compatible point • Stable sampling
ICP Variants • Selecting source points (from one or both meshes) • Matching to points in the other mesh • Weighting the correspondences • Rejecting certain (outlier) point pairs • Assigning an error metric to the current transform • Minimizing the error metric w.r.t. transformation
Closest Compatible Point • Closest point often a bad approximation to corresponding point • Can improve matching effectiveness by restricting match to compatible points • Compatibility of colors [Godin et al. 94] • Compatibility of normals [Pulli 99] • Other possibilities: curvatures, higher-order derivatives, and other local features
ICP Variants • Selecting source points (from one or both meshes) • Matching to points in the other mesh • Weighting the correspondences • Rejecting certain (outlier) point pairs • Assigning an error metric to the current transform • Minimizing the error metric w.r.t. transformation
Selecting Source Points • Use all points • Uniform subsampling • Random sampling • Stable sampling [Gelfand et al. 2003] • Select samples that constrain all degrees of freedomof the rigid-body transformation
Stable Sampling Uniform Sampling Stable Sampling
Covariance Matrix • Aligning transform is given by ATAx = ATb, where • Covariance matrix C = ATA determines the change in error when surfaces are moved from optimal alignment
Sliding Directions • Eigenvectors of C with small eigenvalues correspond to sliding transformations 3 small eigenvalues 2 translation 1 rotation 2 small eigenvalues 1 translation 1 rotation 3 small eigenvalues 3 rotation 1 small eigenvalue 1 translation 1 small eigenvalue 1 rotation [Gelfand]
Stability Analysis Key: 3 DOFs stable 5 DOFs stable 4 DOFs stable 6 DOFs stable
Sample Selection • Select points to prevent small eigenvalues • Based on C obtained from sparse sampling • Simpler variant: normal-space sampling • Select points with uniform distribution of normals • Pro: faster, does not require eigenanalysis • Con: only constrains translation
Result Stability-based or normal-space sampling important for smooth areas with small features Random sampling Normal-space sampling
Selection vs. Weighting • Could achieve same effect with weighting • Hard to ensure enough samples in features except at high sampling rates • However, have to build special data structure • Preprocessing / run-time cost tradeoff
Improving ICP Speed Projection-based matching • Selecting source points (from one or both meshes) • Matching to points in the other mesh • Weighting the correspondences • Rejecting certain (outlier) point pairs • Assigning an error metric to the current transform • Minimizing the error metric w.r.t. transformation
Finding Corresponding Points • Finding closest point is most expensive stageof the ICP algorithm • Brute force search – O(n) • Spatial data structure (e.g., k-d tree) – O(log n)
Projection to Find Correspondences • Idea: use a simpler algorithm to find correspondences • For range images, can simply project point [Blais 95] • Constant-time • Does not require precomputing a spatial data structure
Projection-Based Matching • Slightly worse performance per iteration • Each iteration is one to two orders of magnitude faster than closest-point • Result: can aligntwo range imagesin a few milliseconds,vs. a few seconds
Application • Given: • A scanner that returns range images in real time • Fast ICP • Real-time merging and rendering • Result: 3D model acquisition • Tight feedback loop with user • Can see and fill holes while scanning
Theoretical Analysis of ICP Variants • One way of studying performance is via empirical tests on various scenes • How to analyze performance analytically? • For example, when does point-to-plane help? Under what conditions does projection-based matching work?
f(x) What Does ICP Do? • Two ways of thinking about ICP: • Solving the correspondence problem • Minimizing point-to-surface squared distance • ICP is like Newton’s method on an approximation of the distance function
What Does ICP Do? • Two ways of thinking about ICP: • Solving the correspondence problem • Minimizing point-to-surface squared distance • ICP is like Newton’s method on an approximation of the distance function f’(x)
What Does ICP Do? • Two ways of thinking about ICP: • Solving the correspondence problem • Minimizing point-to-surface squared distance • ICP is like Newton’s method on an approximation of the distance function • ICP variants affect shape ofglobal error function orlocal approximation
Soft Matching and Distance Functions • Soft matching equivalent to standard ICP on (some) filtered surface • Produces filtered version of distance function fewer local minima • Multiresolution minimization [Turk & Levoy 94]or softassign with simulated annealing(good description in [Chui 03])
2D 3D [Mitra et al. 2004] Mitra et al.’s Optimization • Precompute piecewise-quadratic approximation to distance field throughout space • Store in “d2tree” data structure
Mitra et al.’s Optimization • Precompute piecewise-quadratic approximation to distance field throughout space • Store in “d2tree” data structure • At run time, look up quadratic approximants and optimize using Newton’s method • More robust, wider basin of convergence • Often fewer iterations, but more precomputation
Global Registration Goal • Given: n scans around an object • Goal: align them all • First attempt: ICP each scan to one other
Global Registration Goal • Want method for distributing accumulated error among all scans
Approach #1: Avoid the Problem • In some cases, have 1 (possibly low-resolution) scan that covers most of surface • Align all other scans to this “anchor” [Turk 94] • Disadvantage: not always practical to obtain anchor scan
Approach #2: The Greedy Solution • Align each new scan to all previous scans [Masuda 96] • Disadvantages: • Order dependent • Doesn’t spread out error
Approach #3: The Brute-Force Solution • While not converged: • For each scan: • For each point: • For every other scan • Find closest point • Minimize error w.r.t. transforms of all scans • Disadvantage: • Solve (6n)(6n) matrix equation,where n is number of scans